A method for separating and extracting GPR abnormal reflection waves

Abstract

The disclosed embodiments provide a method for separating and extracting GPR anomalous reflection waves, belonging to the field of data processing technology. The method specifically comprises: Step 1, creating a deep Res-CNN network structure; Step 2, establishing a GPR dataset with the deep Res-CNN network structure; Step 3, preprocessing the GPR dataset; Step 4, training a deep Res-CNN network weight model for separating and extracting GPR anomalous reflection waves using the preprocessed GPR dataset; and Step 5, using the deep Res-CNN network weight model to separate anomalous reflection waves from the actual collected data profile and extract valid reflection waves. The disclosed solution enables rapid separation and extraction of anomalous reflection waves from GPR data, laying the foundation for further accurate interpretation of subsurface anomaly features.

Description

Technical Field

[0001] The embodiments of the present disclosure relate to the field of data processing technology, and in particular to a method for separating and extracting GPR abnormal reflection waves. Background Art

[0002] At present, Ground Penetrating Radar (GPR) is a geophysical method that uses high-frequency electromagnetic waves to determine the distribution pattern of substances inside a medium. It is non-destructive, efficient, and convenient, and has been widely used in building quality inspection, road internal disease detection, underground pipeline detection, etc. With the continuous application and popularization of three-dimensional GPR, a large amount of GPR data is often obtained during the GPR data acquisition process in different fields. However, relying solely on experts to analyze and identify abnormal reflection waves in these massive data is not only time-consuming, but also prone to missed detections and false detections. Therefore, it is important to quickly separate and effectively extract reflection waves from massive GPR data, and provide a technology for detecting abnormal reflection waves intelligently, quickly, and efficiently for target separation and extraction in the field of GPR image data processing, which is of great significance for accurately identifying targets.

[0003] GPR technology, based on the underground propagation and scattering characteristics of high-frequency electromagnetic waves, exploits variations in the electromagnetic properties of different materials to detect underground areas. When electromagnetic waves encounter single-point anomalies or columnar scatterers (such as cavities and pipes), they produce reflection hyperbolas in the original GPR profile. Analyzing these anomalous reflection hyperbolas allows for the location and identification of anomalous target features, playing a crucial role in GPR detection. Therefore, extracting these hyperbolic anomaly features is crucial for identifying underground target anomalies.

[0004] Traditional methods for extracting hyperbolas rely on image and signal processing, with professionals identifying the target's reflected waves based on waveform, energy, and other information. Windsor et al. used the generalized Hough transform method to extract GPR reflected waves from buried objects; Terrasse et al. proposed a curvelet transform method to detect hyperbolas in GPR profiles to locate buried pipelines. et al. used second-order statistical parameters to identify and extract features from GPR images to detect pipeline leaks; Zhao et al. combined local planar filters with online dictionary learning to extract small-scale underground targets from GPR diffraction. However, these methods require large storage space and long computation times. Furthermore, due to the complex underground environment and interference from other objects, the separation and extraction of GPR anomalous reflection waves is more difficult and requires more specialized knowledge. Relying solely on manual interpretation is not only inefficient but also prone to missed detections and false detections. In recent years, deep learning technology has been used in the remote sensing field to process remote sensing images of buildings, roads, and other objects, providing effective information for urban planning and disaster monitoring. Based on this, deep learning technology has gradually attracted the interest of researchers in the GPR field, bringing new ideas to the detection and separation of buried objects. Lameri et al. proposed a mine detection method based on convolutional neural networks to identify mine reflection curves in GPR profiles. Picetti et al. applied convolutional autoencoders to B-scans for mine detection, providing a new approach for GPR image detection. Zhang et al. proposed a deep learning framework based on generative adversarial networks to detect hyperbolic reflections through an end-to-end solution. Hou et al. improved the masked region-based convolutional network architecture, achieving good accuracy in detecting and segmenting abnormal features in GPR images. Although deep learning networks have demonstrated excellent performance in detecting hyperbolas in GPR images, separating and extracting anomalous reflections from noisy GPR profiles remains a challenging task. In GPR image interpretation, separating and extracting anomalous reflections is a key step in obtaining target information from GPR profiles. Relying solely on manual identification of hyperbolic reflection features in massive amounts of data is time-consuming and labor-intensive.

[0005] It can be seen that there is an urgent need for an efficient and accurate method for separating and extracting GPR abnormal reflection waves. Summary of the Invention

[0006] In view of this, an embodiment of the present disclosure provides a method for separating and extracting GPR abnormal reflection waves, which at least partially solves the problems of poor extraction efficiency and accuracy in the prior art.

[0007] The present disclosure provides a method for separating and extracting abnormal GPR reflection waves, including:

[0008] Step 1: Create a deep Res-CNN network structure;

[0009] Step 2: Establish a GPR dataset with a deep Res-CNN network structure;

[0010] Step 3, preprocessing the GPR data set;

[0011] Step 4: Use the preprocessed GPR dataset to train a deep Res-CNN network weight model for GPR abnormal reflection wave separation and extraction;

[0012] Step 5: Use the deep Res-CNN network weight model to separate the abnormal reflection waves in the actual collected data profile and extract the effective reflection waves.

[0013] According to a specific implementation method of an embodiment of the present disclosure, the network structure of the deep Res-CNN includes 5 downsampling blocks, 5 upsampling blocks and 1 output block, and a residual block composed of a Skip Connection structure is arranged between the downsampling block and the upsampling block. Each downsampling block includes a convolution layer, a LeakyReLU activation function layer and a MaxPooling layer. The convolution layer includes 4 2×2 and 1 1×1 convolution kernels in sequence, and the input channels are 16, 32, 64, 128, and 256 in sequence. Each upsampling block includes a transposed convolution layer, a Batch Normalization layer and a ReLU activation function layer. The convolution layer includes 1 1×1 and 4 2×2 convolution kernels in sequence, and the input channels are 128, 64, 32, 16, and 16 in sequence. The output layer includes a transposed convolution layer and a Sigmoid activation function layer.

[0014] According to a specific implementation of the embodiment of the present disclosure, step 2 specifically includes:

[0015] A GPR forward modeling program is used to establish anomaly models in a uniform medium and a random medium thereof, and a forward modeling data set is obtained as the GPR data set.

[0016] According to a specific implementation of the embodiment of the present disclosure, step 3 specifically includes:

[0017] The robust non-negative matrix factorization method is used to remove the direct waves in the GPR dataset and retain the hyperbolic characteristics of the underground abnormal reflection waves.

[0018] According to a specific implementation of the embodiment of the present disclosure, step 4 specifically includes:

[0019] Step 4.1: randomly divide the preprocessed GPR dataset into training set, validation set and test set according to the preset ratio;

[0020] Step 4.2: In the downsampling block, the LeakyReLU activation function is used to prevent the loss of neuron information caused by negative values in the downsampling block. In the upsampling block, no negative parameters are involved, and the ReLU activation function is directly used to restore the data. In addition, the sigmoid activation function is used to restore the output mask data to the interval [0, 1].

[0021] Step 4.3, introduce the residual layer of the Skip Connection structure;

[0022] Step 4.4: Add a batch normalization layer during the training of the deep Res-CNN network structure, and keep the input of each layer of the deep Res-CNN network with the same distribution;

[0023] In step 4.5, we select the binary cross entropy function as the loss function to evaluate the quality of the training deep Res-CNN network structure.

[0024] In step 4.6, the RMSprop method with adaptive learning rate is selected to iteratively learn each network training parameter. The update iteration process is as follows:

[0025]

[0026]

[0027] In the formula, E[g 2 ] is the moving average of the squared gradient, θ is the parameter to be updated, g t is the gradient of θ at time t, η is the learning rate, and ∈ is the minimum parameter to prevent the denominator from being 0;

[0028] In step 4.7, set the single batch training of multiple samples of the deep Res-CNN network structure, the preset learning rate and the number of network training times, and obtain the trained deep Res-CNN network weight model when the loss function training is stable.

[0029] According to a specific implementation of the embodiment of the present disclosure, the binary cross entropy function is expressed as:

[0030]

[0031] Where y i is the true binary label value, i=1 represents the positive class value, i=0 represents the negative class value, p(y i ) is the predicted probability value.

[0032] According to a specific implementation of the embodiment of the present disclosure, step 5 specifically includes:

[0033] The actual collected data profile is input into the trained deep Res-CNN network weight model to separate the abnormal reflection waves to obtain mask data, and the mask data is applied to the original GPR profile to extract the effective reflection waves.

[0034] The GPR abnormal reflection wave separation and extraction scheme in the embodiment of the present disclosure includes: step 1, creating a deep Res-CNN network structure; step 2, establishing a GPR data set of the deep Res-CNN network structure; step 3, preprocessing the GPR data set; step 4, training a deep Res-CNN network weight model for GPR abnormal reflection wave separation and extraction using the preprocessed GPR data set; step 5, using the deep Res-CNN network weight model to separate abnormal reflection waves in the actually collected data profile and extract effective reflection waves.

[0035] The beneficial effects of the embodiments of the present disclosure are as follows: through the scheme of the present disclosure, a deep convolutional network structure with a skip connection layer is created. This structure can overcome the problem of gradient explosion or vanishing caused by the increase in the number of network layers, thereby realizing the separation and extraction of GPR abnormal reflection waves. In the established training data set, the network structure is trained using forward simulation data under uniform media and its random media to solve the problem of training data volume; in addition, a robust non-negative matrix decomposition method is used as a preprocessing strategy for the training data set to reduce the impact of the strong energy of the direct wave on the abnormal reflection wave. Finally, the trained deep Res-CNN network weight model is applied to the forward simulation data profile and the actual collected data profile, indicating that the network structure constructed by the present invention has good applicability and generalization ability in different GPR data profiles, and can quickly separate and extract abnormal reflection waves in GPR data, laying the foundation for further accurate interpretation of underground abnormal characteristics. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the technical solutions of the embodiments of the present disclosure, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present disclosure. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0037] Figure 1 A schematic flow chart of a method for separating and extracting GPR abnormal reflection waves provided in an embodiment of the present disclosure;

[0038] Figure 2 A schematic diagram of a deep Res-CNN network structure provided in an embodiment of the present disclosure;

[0039] Figure 3 A comparison diagram before and after removing direct waves using a robust non-negative matrix decomposition method provided by an embodiment of the present disclosure;

[0040] Figure 4 A loss function curve diagram during a network training process provided by an embodiment of the present disclosure;

[0041] Figure 5 A schematic cross-sectional diagram of a forward model and simulation result data provided by an embodiment of the present disclosure;

[0042] Figure 6 A diagram comparing the separation and extraction results of GPR abnormal reflection waves using a deep Res-CNN network structure under forward simulation data provided by an embodiment of the present disclosure with the traditional OTSU threshold segmentation algorithm and the iterative threshold segmentation algorithm;

[0043] Figure 7 A schematic diagram of measured data results provided by an embodiment of the present disclosure;

[0044] Figure 8 This is a schematic diagram showing the comparison between the GPR abnormal reflection wave separation and extraction results of a deep Res-CNN network structure under measured data provided by an embodiment of the present disclosure and the traditional OTSU threshold segmentation algorithm and the iterative threshold segmentation algorithm. DETAILED DESCRIPTION

[0045] The embodiments of the present disclosure are described in detail below with reference to the accompanying drawings.

[0046] The following describes the embodiments of the present disclosure through specific examples, and those skilled in the art can easily understand other advantages and effects of the present disclosure from the contents disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present disclosure, rather than all of the embodiments. The present disclosure can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present disclosure. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Based on the embodiments in the present disclosure, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present disclosure.

[0047] It should be noted that various aspects of the embodiments within the scope of the appended claims are described below. It should be apparent that the aspects described herein can be embodied in a wide variety of forms, and any specific structure and / or function described herein is merely illustrative. Based on this disclosure, it should be understood by those skilled in the art that an aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects described herein can be used to implement an apparatus and / or practice a method. In addition, other structures and / or functionalities other than one or more of the aspects described herein can be used to implement this apparatus and / or practice this method.

[0048] It should also be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present disclosure. The illustrations only show components related to the present disclosure and are not drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component can be changed at will, and the component layout type may also be more complicated.

[0049] Additionally, in the following description, specific details are provided to provide a thorough understanding of the examples. However, one skilled in the art will appreciate that the aspects described can be practiced without these specific details.

[0050] The embodiments of the present disclosure provide a method for separating and extracting GPR abnormal reflection waves, which can be applied to the separation and extraction process of abnormal reflection waves from ground penetrating radar in scenarios such as building quality inspection, road internal disease inspection, and underground pipeline detection.

[0051] See also Figure 1 , is a flow chart of a method for separating and extracting abnormal GPR reflection waves provided by an embodiment of the present disclosure. Figure 1 As shown, the method mainly includes the following steps:

[0052] Step 1: Create a deep Res-CNN network structure;

[0053] Optionally, the network structure of the deep Res-CNN includes 5 downsampling blocks, 5 upsampling blocks and 1 output block, and a residual block composed of a Skip Connection structure is arranged between the downsampling blocks and the upsampling blocks. Each downsampling block includes a convolution layer, a LeakyReLU activation function layer and a MaxPooling layer. The convolution layer includes 4 2×2 and 1 1×1 convolution kernels in sequence, and the input channels are 16, 32, 64, 128, and 256 in sequence. Each upsampling block includes a transposed convolution layer, a BatchNormalization layer and a ReLU activation function layer. The convolution layer includes 1 1×1 and 4 2×2 convolution kernels in sequence, and the input channels are 128, 64, 32, 16, and 16 in sequence. The output layer includes a transposed convolution layer and a Sigmoid activation function layer.

[0054] For example, creating Figure 2 The deep Res-CNN network structure shown in the figure can quickly separate and extract abnormal reflection waves. The specific description of the deep Res-CNN network structure is as follows:

[0055] The deep Res-CNN network structure is composed of 5 downsampling blocks, 5 upsampling blocks and 1 output block. The residual block composed of the Skip Connection structure between the downsampling blocks and the upsampling blocks overcomes the problem of gradient explosion or disappearance caused by the increase in the number of layers in the network structure, so that the upper layer information is fully input into the lower layer network. In each downsampling block, it consists of a convolution layer, a LeakyReLU activation function layer, and a MaxPooling layer. The convolution layer uses four 2×2 and one 1×1 convolution kernels to perform two-dimensional convolution operations in sequence. The input channels are 16, 32, 64, 128, and 256, respectively, to achieve downsampling and dimensionality reduction of the input image. In each upsampling block, it consists of a transposed convolution layer, a Batch Normalization layer, and a ReLU activation function layer. The convolution layer uses one 1×1 and four 2×2 convolution kernels to perform two-dimensional convolution operations in sequence. The input channels are 128, 64, 32, 16, and 16, respectively, to achieve dimensionality increase of the upsampled image. Finally, the transposed convolution layer and Sigmoid activation function layer of the output layer generate mask data, and the abnormal reflection wave is separated and extracted.

[0056] Step 2: Establish a GPR dataset with a deep Res-CNN network structure;

[0057] Furthermore, the step 2 specifically includes:

[0058] A GPR forward modeling program is used to establish anomaly models in a uniform medium and a random medium thereof, and a forward modeling data set is obtained as the GPR data set.

[0059] In specific implementation, the GPR forward modeling program can be used to establish anomaly models under uniform media and its random media to obtain a large number of forward modeling data sets. The random medium model can produce clutter profiles generated by background medium factors, and the uniform medium model can produce homogeneous profiles, and mask data can be obtained at the same time.

[0060] Step 3, preprocessing the GPR data set;

[0061] Furthermore, the step 3 specifically includes:

[0062] The robust non-negative matrix factorization method is used to remove the direct waves in the GPR dataset and retain the hyperbolic characteristics of the underground abnormal reflection waves.

[0063] In specific implementation, the robust non-negative matrix decomposition method can be used to perform direct wave removal preprocessing in the obtained GPR training data set, which can effectively remove the direct wave and retain the hyperbolic characteristics of the underground abnormal reflection wave, such as Figure 3As shown in the figure, (a) is before the direct wave is removed, and (b) is after the direct wave is removed. Since the direct wave energy is very strong, the effective underground reflection wave is easily submerged and difficult to distinguish clearly. Therefore, reducing the influence of the strong energy of the direct wave on the abnormal reflection is an important step in realizing the separation and extraction of anomalies in GPR signals in the present invention.

[0064] Step 4: Use the preprocessed GPR dataset to train a deep Res-CNN network weight model for GPR abnormal reflection wave separation and extraction;

[0065] Based on the above embodiment, step 4 specifically includes:

[0066] Step 4.1: randomly divide the preprocessed GPR dataset into training set, validation set and test set according to the preset ratio;

[0067] Step 4.2: In the downsampling block, the LeakyReLU activation function is used to prevent the loss of neuron information caused by negative values in the downsampling block. In the upsampling block, no negative parameters are involved, and the ReLU activation function is directly used to restore the data. In addition, the sigmoid activation function is used to restore the output mask data to the interval [0, 1].

[0068] Step 4.3, introduce the residual layer of the Skip Connection structure;

[0069] Step 4.4: Add a batch normalization layer during the training of the deep Res-CNN network structure, and keep the input of each layer of the deep Res-CNN network with the same distribution;

[0070] In step 4.5, we select the binary cross entropy function as the loss function to evaluate the quality of the training deep Res-CNN network structure.

[0071] In step 4.6, the RMSprop method with adaptive learning rate is selected to iteratively learn each network training parameter. The update iteration process is as follows:

[0072]

[0073]

[0074] In the formula, E[g 2 ] is the moving average of the squared gradient, θ is the parameter to be updated, g t is the gradient of θ at time t, η is the learning rate, and ∈ is the minimum parameter to prevent the denominator from being 0;

[0075] In step 4.7, set the single batch training of multiple samples of the deep Res-CNN network structure, the preset learning rate and the number of network training times, and obtain the trained deep Res-CNN network weight model when the loss function training is stable.

[0076] Furthermore, the binary cross entropy function is expressed as:

[0077]

[0078] Where y i is the true binary label value, i=1 represents the positive class value, i=0 represents the negative class value, p(y i ) is the predicted probability value.

[0079] In the specific implementation, the deep Res-CNN network weight model for GPR abnormal reflection wave separation and extraction is trained. The specific steps are as follows:

[0080] The specific steps are as follows:

[0081] A. Randomly divide the preprocessed GPR dataset into training set, validation set and test set according to the ratio of 8:1:1;

[0082] B. In the downsampling block, the LeakyReLU activation function is used to prevent the degradation problem of neuronal information loss caused by negative values in the downsampling block; in the upsampling block, no negative parameters are involved, and the ReLU activation function is directly used to restore the data; finally, the sigmoid activation function is used to ensure that the output mask data is restored to the interval [0,1].

[0083] C. Introducing the residual layer of the Skip Connection structure prevents the loss of upper-layer information from causing the lower-layer network to be unable to recover, as well as the gradient vanishing problem caused by the increase in the number of layers, ensuring that the upper-layer information can be fully input into the lower layer of the network.

[0084] D. Add a batch normalization layer during the training of the deep Res-CNN network structure to overcome the slow convergence problem caused by discontinuous data distribution. At the same time, the input of each layer of the neural network maintains the same distribution.

[0085] E. Choose to use the binary cross entropy function as the loss function to evaluate the quality of the training network structure. The binary cross entropy function can be expressed as:

[0086]

[0087] Where y i is the true binary label value, i=1 represents the positive class value, i=0 represents the negative class value, p(y i ) is the predicted probability value.

[0088] F. Select the RMSprop method with adaptive learning rate to learn each network training parameter. The update iterative process is as follows:

[0089]

[0090]

[0091] In the formula, E[g 2 ] is the moving average of the squared gradient, θ is the parameter to be updated, g t is the gradient of θ at time t, η is the learning rate, and ∈ is the minimum parameter to prevent the denominator from being 0.

[0092] G. Set the single batch training of the deep Res-CNN network structure to 32 samples, the learning rate η is 0.001, and the number of network training times is set to 600 times. The loss function iterative convergence curve is as follows Figure 4 As shown in the figure, we can see that the loss function gradually stabilizes after training iterations, without overfitting, and finally a trained weight model is obtained.

[0093] Step 5: Use the deep Res-CNN network weight model to separate the abnormal reflection waves in the actual collected data profile and extract the effective reflection waves.

[0094] Based on the above embodiment, step 5 specifically includes:

[0095] The actual collected data profile is input into the trained deep Res-CNN network weight model to separate the abnormal reflection waves to obtain mask data, and the mask data is applied to the original GPR profile to extract the effective reflection waves.

[0096] In specific implementation, after training the deep Res-CNN network weight model, the deep Res-CNN network weight model can be used to separate abnormal reflection waves in the actual collected data profile and extract valid reflection waves. At the same time, in order to illustrate the difference between this scheme and the traditional method, the separation and extraction results of the deep Res-CNN network can also be compared with the effects of the traditional OTSU threshold segmentation algorithm and iterative threshold segmentation algorithm.

[0097] The specific process is as follows:

[0098] A. Figure 5As shown, the forward model shown in (a) is established. The model is a random medium model. The first layer is an air layer, the second layer has a relative dielectric constant with a mean of 4 and a variance of 0.1, and the third layer has a relative dielectric constant with a mean of 9 and a variance of 0.1. Circular anomalies are used to simulate pipelines of different materials, and irregular bodies are used to simulate cavities. The simulation area is simulated using a 400MHz center frequency antenna with a spatial step of 0.02m, a time step of 0.04ns, and a time window of 80ns. The trained deep Res-CNN network weight model is applied to the forward simulation profile shown in (b), and the separation and extraction results are compared with the traditional OTSU threshold segmentation algorithm and the iterative threshold segmentation algorithm, as shown in FIG. Figure 6 Figure 2 shows the results of GPR anomaly reflection wave separation and extraction using the traditional OTSU threshold segmentation algorithm for forward modeling data, (c) and (d) using the traditional iterative threshold segmentation algorithm for forward modeling data, and (e) and (f) using the deep Res-CNN network structure for forward modeling data. As shown in (e) and (f), the deep Res-CNN network structure can better remove the clutter caused by the background medium, separate and extract clear target reflection waves, and locate the anomalous reflections. This demonstrates that the deep Res-CNN network structure created by the present invention has good applicability in GPR data profiles.

[0099] B. To further demonstrate that the created deep Res-CNN network structure has a certain generalization ability, the actual collected GPR data, such as Figure 7 As shown in , after preprocessing, it is input into the trained deep Res-CNN network weight model and compared with the traditional OTSU threshold segmentation algorithm and iterative threshold segmentation algorithm, as shown in Figure 8 (a) and (b) are the results of GPR abnormal reflection wave separation and extraction using the traditional OTSU threshold segmentation algorithm under measured data, (c) and (d) are the results of GPR abnormal reflection wave separation and extraction using the traditional iterative threshold segmentation algorithm under measured data, and (e) and (f) are the results of GPR abnormal reflection wave separation and extraction using the deep Res-CNN network structure under measured data. Figure 8 (e) and (f) show that the deep Res-CNN network structure created by the present invention can effectively suppress the influence of background clutter on the abnormal reflection waves of the actual GPR profile, and can quickly separate and extract clearer abnormal reflection waves, verifying the generalization ability of the network structure of the present invention.

[0100] It should be understood that various parts of the present disclosure can be implemented in hardware, software, firmware, or a combination thereof.

[0101] The above description is merely a specific embodiment of the present disclosure, but the scope of protection of the present disclosure is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this disclosure should be included in the scope of protection of the present disclosure. Therefore, the scope of protection of the present disclosure should be based on the scope of protection of the claims.

Claims

1. A method for separating and extracting abnormal GPR reflection waves, characterized in that: include: Step 1, create a deep Res-CNN network structure, wherein the deep Res-CNN network structure includes 5 downsampling blocks, 5 upsampling blocks and 1 output block, a residual block composed of a SkipConnection structure is set between the downsampling block and the upsampling block, each downsampling block includes a convolution layer, a LeakyReLU activation function layer and a MaxPooling layer, the convolution layer includes 4 2×2 and 1 1×1 convolution kernels in sequence, and the input channels are 16, 32, 64, 128, and 256 in sequence, each upsampling block includes a transposed convolution layer, a BatchNormalization layer and a ReLU activation function layer, the convolution layer includes 1 1×1 and 4 2×2 convolution kernels in sequence, the input channels are 128, 64, 32, 16, and 16 in sequence, and the output layer includes a transposed convolution layer and a Sigmoid activation function layer; Step 2: Establish a GPR dataset with a deep Res-CNN network structure; Step 3, preprocessing the GPR data set; Step 4: Use the preprocessed GPR dataset to train a deep Res-CNN network weight model for GPR abnormal reflection wave separation and extraction; The step 4 specifically includes: Step 4.1: randomly divide the preprocessed GPR dataset into training set, validation set and test set according to the preset ratio; Step 4.2: In the downsampling block, the LeakyReLU activation function is used to prevent the loss of neuron information caused by negative values in the downsampling block. In the upsampling block, no negative parameters are involved, and the ReLU activation function is directly used to restore the data. In addition, the sigmoid activation function is used to restore the output mask data to the interval [0, 1]. Step 4.3, introduce the residual layer of SkipConnection structure; Step 4.4: Add a batch normalization layer during the training of the deep Res-CNN network structure, and keep the input of each layer of the deep Res-CNN network with the same distribution; In step 4.5, the binary cross entropy function is selected as the loss function to evaluate the quality of the training deep Res-CNN network structure, where the binary cross entropy function is expressed as: Where y i is the true binary label value, i=1 represents the positive class value, i=0 represents the negative class value, p(y i ) is the predicted probability value; In step 4.6, the RMSprop method with adaptive learning rate is selected to iteratively learn each network training parameter. The update iteration process is as follows: In the formula, E[g 2 ] is the moving average of the squared gradient, θ is the parameter to be updated, g t is the gradient of θ at time t, η is the learning rate, and ∈ is the minimum parameter to prevent the denominator from being 0; Step 4.7: Set the single batch training of multiple samples of the deep Res-CNN network structure, the preset learning rate and the number of network training times. When the loss function training is stable, the trained deep Res-CNN network weight model is obtained. Step 5: Use the deep Res-CNN network weight model to separate the abnormal reflection waves in the actual collected data profile and extract the effective reflection waves.

2. The method according to claim 1, characterized in that , the step 2 specifically includes: A GPR forward modeling program is used to establish anomaly models in a uniform medium and a random medium thereof, and a forward modeling data set is obtained as the GPR data set.

3. The method according to claim 1, characterized in that , the step 3 specifically includes: The robust non-negative matrix factorization method is used to remove the direct waves in the GPR dataset and retain the hyperbolic characteristics of the underground abnormal reflection waves.

4. The method according to claim 1, characterized in that , the step 5 specifically includes: The actual collected data profile is input into the trained deep Res-CNN network weight model to separate the abnormal reflection waves to obtain mask data, and the mask data is applied to the original GPR profile to extract the effective reflection waves.

Images